S2298
Physics - Machine learning and AI algorithms
ESTRO 2026
Results: AUC, accuracy, precision and recall of MLMs were, respectively, in the following ranges: 0.71-0.78, 0.65- 0.74, 0.49-0.89 and 0.49-0.80. APP used prospectively predicted always planned outcomes successfully. Paired t-tests showed significative differences between CTV and PTV coverage with the two optimization workflow; single tail t-test assessed best results with the new one in case of predicted failure. Preliminary results from the CNN showed accuracy 0.65±0.02, AUC 0.77±0.06, precision 0.55 ± 0.05, and recall 1.00 ± 0.03. The activation maps showed high activation around CTV and OARs.
Classification with Deep Convolutional Neural Networks. NIPS 2012. Chen S, et al. Med3D: Transfer Learning for 3D Medical Image Analysis. arXiv 2019. Selvaraju R, et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. ICCV 2017. Keywords: Autoplanning, prostate, optimization Digital Poster Highlight 2808 Enhancing deep learning NTCP prediction for late xerostomia using mid-treatment CT scans Luuk van der Hoek, Hendrike Neh, Daniel C. MacRae, Suzanne P.M. de Vette, Nanna M. Sijtsema, Johannes A. Langendijk, Peter M.A. van Ooijen, Lisanne V. van Dijk Radiotherapy, University Medical Center Groningen (UMCG), Groningen, Netherlands Purpose/Objective NTCP models estimate the risk of treatment-related side effects and can inform clinical decision making in radiotherapy, including the selection of optimal treatment plans. As adaptive radiotherapy becomes part of standard clinical practice, integrating NTCP predictions in adaptive strategies might further empower personalized treatment. Mid-treatment imaging may enhance NTCP model predictions by incorporating changes occurring during the course of radiotherapy. In this study, we aimed to improve a deep learning-based NTCP model for xerostomia by incorporating mid-treatment CT data for head and neck cancer (HNC) patients. Material/Methods Data was collected from HNC patients receiving primary radiotherapy between 2014 and 2024. Patient-rated xerostomia was assessed using the EORTC QLQ-H&N35 questionnaire. The primary endpoint was moderate-to-severe xerostomia at 12 months after treatment. Collected data included pre-treatment CTs, nominal dose, repeat CTs acquired during treatment, clinical features (age, sex, and acute xerostomia scores). Mid- treatment CTs were acquired at approximately week 4 of treatment, corresponding to 15 to 24 delivered fractions depending on whether patients received conventional or accelerated radiotherapy. Two models were trained to demonstrate the effect of incorporating mid-treatment data. First a pre- treatment model using only pre-treatment data as input: pre-treatment CT, nominal dose, and pre- treatment clinical features (baseline xerostomia scores, age, and sex). Secondly a mid-treatment model which extended the input to include the mid- treatment CT and acute xerostomia scores (collected weekly at weeks 2–4). Both models were ResNet-based
Conclusion: MLMs showed limited performance due to the small training and test datasets. Integrating them into the APP using a bagging strategy allowed to leverage each model’s strengths and achieve accurate predictions. The new optimization workflow improved target coverage while maintaining OAR constraints. Preliminary CNN results showed high sensitivity but limited precision. This conservative behaviour may be advantageous in clinical use, where overestimating failures favours a more efficient workflow. Additional validation on larger datasets is necessary before the CNN can replace traditional MLMs and enable fully automated, human-independent prediction. References: Tsang DS, et al. A Prospective Study of Machine Learning-Assisted Radiation Therapy Planning for Patients Receiving 54 Gy to the Brain. Int J Radiat Oncol Biol Phys 2024. Kugel C, et al. Development and Clinical Implementation of Auto-Planning in Pinnacle for Head and Neck Cancer. Strahlentherapie und Onkologie 2019. Krizhevsky A, et al. ImageNet
Made with FlippingBook - Share PDF online